ECTI Transactions on Computer and Information Technology (ECTI-CIT)

Improving Course Recommendations: An RBM-Based Explainable Framework for Student Performance in E-Learning

2026-04-23T09:09:39+07:00

The vast expansion of online learning has increased the need for intelligent systems to guide learners toward suitable learning paths. Although educational recommender systems address this challenge by personalizing recommendations, many existing approaches focus primarily on accuracy and provide limited insight into why recommendations are made. This study proposes EduExplain, an explainable collaborative filtering framework based on a Restricted Boltzmann Machine (RBM) that combines strong predictive performance with neighbor-similarity explanations to generate interpretable recommendations. The framework integrates a prediction engine with an explanation module that produces natural-language justifications derived from the performance of similar learners. We evaluated the framework using a multi-domain experimental design on the OULAD educational dataset and the Goodreads dataset, comparing it against four established baselines. The results show that EduExplain achieves a 20% reduction in prediction error over traditional collaborative filtering (RMSE: 0.873 on OULAD) and delivers high-quality, interpretable explanations (coherence: 0.89), demonstrating that accuracy and explainability can coexist in e-learning systems. Unlike prior approaches that rely on external review data or treat explainability as a post-hoc step, EduExplain embeds explanation generation directly into the RBM pipeline using only interaction data. An ablation study confirmed each component's necessity for both accuracy and interpretability. These findings demonstrate that explainable recommender systems can enhance decision-making in e-learning environments.

D-CAD: Decentralized Continual Anomaly Detection through Collaborative Knowledge Fusion in Wireless Sensor Networks

2026-04-09T08:38:05+07:00

Wireless Sensor Networks (WSNs) are critical for real-time monitoring in industrial and environmental applications, where robust anomaly detection is essential for safety and efficiency. However, existing approaches face a trilemma: centralized methods create communication bottlenecks and single points of failure, isolated on-device learning cannot leverage collective knowledge, and static models degrade under real-world concept drifts. To overcome these limitations, we propose D-CAD, a novel framework for Decentralized Continual Anomaly Detection. D-CAD enables sensor nodes to collaboratively learn and adapt detection models over time using a lightweight, gossip-based knowledge fusion protocol, thereby eliminating the need for a central coordinator. Our method combines local lightweight autoencoders with a replay-based memory buffer to mitigate catastrophic forgetting and a dynamic weighting mechanism for effective peer-to-peer model fusion. Evaluated on the SWaT industrial dataset and a simulated non-IID WSN testbed, D-CAD achieved an average F1-score of 0.92, outperforming a centralized continual learner by 8% and isolated nodes by 23%. It maintains high accuracy while reducing the total network communication overhead by 65% compared with standard Federated Learning. Thus, D-CAD provides a scalable, robust, and communication-efficient paradigm for lifelong anomaly detection in distributed-sensing systems.

Improving Hard Negative Handling for Individual Dog Re-identication with Facial Biometrics

2026-04-22T10:42:49+07:00

Visual biometrics, such as facial recognition, play an important role in authentication and identification. This concept extends to animal biometrics, particularly for identifying individual dogs, which is useful for veterinary care, ownership verification, population control, and disease monitoring. However, distinguishing visually similar dogs within the same breed remains challenging due to hard negative samples. In this work, we propose a deep learning model for individual dog re-identification using contrastive learning with a binary cross-entropy loss function. Instead of treating each dog as a separate class, the proposed framework learns a similarity function from positive and negative image pairs, including a high proportion of hard negative pairs during training, to better capture fine-grained facial differences. Experimental results show that the proposed model achieves 96.83% accuracy on hard negative samples and 81.48% accuracy on the public Flickr dog dataset in a zero-shot setting, demonstrating improved performance in same-breed dog re-identification over traditional multiclass classification approaches.

A Vision Transformer-Enhanced ResU-Net50 Model for Accurate MRI Brain Tumor Segmentation Using DBO Optimization

2026-04-02T20:08:41+07:00

For precise evaluation and therapeutic planning, automatic brain cancer segmentation from MRI scans is essential. In this work, a hybrid deep learning framework for image segmentation comprising a Vision transformer and ResU-Net50 is proposed. It combines the strong local feature extraction capabilities of ResNet50 with the global representation learning capabilities of a Vision Transformer embedded in a U-Net framework. To improve mask quality and prediction stability, a preprocessing pipeline powered by Dung Beetle Optimization (DBO) is used to optimize thresholding, morphological enhancement, and test-time augmentation. Comparing DBO-based preprocessing to non-optimized preprocessing, experimental results on a T1-CE MRI dataset with 3064 annotated images show that the former greatly enhances segmentation performance, resulting in increases of 2.3% in precision, 1.9% in Dice Coefficient, and 3.2% in Jaccard Index. With Dice scores of 0.9630, 0.9873, and 0.9862 for pituitary, glioma, and meningioma tumors, respectively, the suggested Vision-ResU- Net50 model further achieves exceptional tumor-wise segmentation results, outperforming cutting-edge techniques like Edge U-Net, Improved U-Net, CNN, and deep learning models by margins of 417% in Dice and 515% in Jaccard Index, the suggested approach is a highly efficient and dependable for clinical brain tumor analysis.

A Comparative Study of Machine Learning and Deep Learning Approaches for Handwritten Sanskrit Recognition

2026-03-27T12:31:43+07:00

Sanskrit holds immense historical, cultural, and scientific value, with many handwritten manuscripts preserving knowledge in philosophy and traditional medicine like Ayurveda. It is essential to digitize these materials to preserve this legacy and provide broader access. Handwritten Sanskrit recognition remains an underexplored yet essential area within Optical Character Recognition (OCR), largely due to the complexity of the script. Elements such as the Shirolekha (headline), compound characters, modifiers, and irregular character boundaries contribute to the difficulty of accurate recognition. This paper presents a comprehensive study of classical and contemporary techniques adopted for Sanskrit and Devanagari script recognition. The study explores machine learning methods such as Support Vector Machines (SVM), k-Nearest Neighbors (KNN), and Hidden Markov Models (HMM), alongside modern deep learning models including CNNs, LSTMs, Bidirectional LSTMs, and CapsNet. The analysis shows that while traditional methods work well with limited data, deep learning approaches achieve higher accuracy, like CNN-based architectures, which reach up to 99.65% but require substantially larger datasets and risk overfitting. A key finding is that most existing research focuses on isolated characters, leaving word-level recognition and complex conjuncts largely unaddressed. Persistent issues such as image noise, overlapping characters, and the lack of large-scale Sanskrit word datasets continue to hinder progress. This study highlights existing gaps and proposes future directions, emphasizing the need for hybrid deep learning models, linguistic context via NLP, and benchmark datasets for improved OCR systems.

Automated Dehydration Assessment in Cannabis via YOLOv8 Localization and Multi-Modal Feature Classification

2026-04-22T10:39:55+07:00

Water management is a fundamental aspect of plant physiology, directly influencing photosynthesis and nutrient intake. The objective of this study is to develop and validate a non-invasive, automated pipeline for quantifying water stress in Cannabis sativa L. using computer vision and physiological ground-truth measurements. The study utilized a dataset of 988 original outdoor images, which were expanded to 4,940 images through data augmentation to ensure model robustness. The process integrates the state-of-the-art YOLOv8 model for leaf localization, achieving a mean Average Precision (mAP) of 98.3 percent. A multi-modal feature set was constructed by combining color statistics derived from four color space transformations (RGB, HSV, LAB, and YCrCb) and Gray-Level Co-occurrence Matrix (GLCM) texture descriptors. The features were used to train a classifier to categorize dehydration into six severity levels (Normal, Mild, Moderate, Distinct, Severe, and Extreme), and the classifier was verified against ground-truth leaf water potential (Ψleaf) measurements. The nd- ings show that a Cubic Support Vector Machine (SVM) outperformed Artificial Neural Networks, achieving an overall accuracy of 86.2%. The research provides a scalable, high-precision solution for real-time monitoring in precision agriculture.

Benchmarking Faster R-CNN Backbones for Explainable Lemon Leaf Disease Detection

2026-06-12T13:50:02+07:00

Accurate and timely detection of lemon leaf diseases is important for maintaining yield and fruit quality, yet manual scouting is labor intensive and subjective. This study explores a region-based deep learning method for symptom localization by benchmarking Faster R-CNN with various backbone feature extractors, including ResNet50, ResNet101, VGG16, and MobileNetV3, on a dataset of lemon leaf diseases annotated with bounding boxes. To ensure a fair comparison, all models are trained using identical preprocessing, augmentation, and optimization settings. They are evaluated based on detection metrics at multiple Intersection over Union (IoU) thresholds, which include mean Average Precision (mAP) at 0.5 to 0.95, mAP at 0.50, mAP at 0.75, average recall, and mean IoU. Soft Non-Maximum Suppression is also applied during inference to reduce missed detections when lesions appear close together in clusters. Results indicate that Faster R-CNN with a ResNet50 backbone and Soft Non-Maximum Suppression provides the best overall performance, achieving an mAP from 0.5 to 0.95 of 0.6584 and an mAP at 0.75 of 0.8052, while maintaining a strong recall of 0.7326 and high localization quality with a mean IoU of 0.8203. To ensure trustworthy deployment, model explanations are generated using Grad-CAM and LIME. These methods demonstrate that the detector primarily focuses on symptomatic regions instead of background patterns. Rather than proposing a new detector architecture, this work provides a controlled benchmark of backbone selection under a unified Faster R-CNN pipeline. The proposed pipeline presents an effective and interpretable solution for monitoring lemon leaf disease within the evaluated dataset and experimental setting. It also offers practical guidance for selecting backbones in region-based agricultural inspection systems.

Mining User Mobility Insights from Public Wi-Fi Data Using Association Rules in Urban Riverfront Areas

2026-05-21T05:19:20+07:00

Understanding user mobility patterns is essential for effective urban planning and resource management. This study employs Association Rule Mining to analyze public Wi-Fi data collected from 11 access points along the Mekong River in Sri Chiang Mai, Thailand, from May 2022 to December 2023. By examining co-occurring movement patterns in over 73.7 million connection records, the research uncovers key insights into human behavior in the area. The findings highlight the area in front of the Fresh Market as a central destination, with confidence values exceeding 0.99 for related movement rules. The results reveal pronounced temporal variations in movement patterns, with transitions from commercial areas in the mornings to leisure spaces in the afternoons and evenings. Weekday patterns differ notably from weekend behaviors, reflecting how time influences urban space utilization. These insights provide urban planners and policymakers with data-driven evidence to optimize infrastructure development, enhance public spaces, and improve resource allocation. Although this study is limited to Wi-Fi data, it provides significant contributions to the development of smart cities that are more responsive and sustainable, paving the way for improved urban living experiences and more efficient resource management. Future work could integrate multiple data sources to enable more comprehensive mobility analysis and advance sustainable urban development goals.

An AI-based Mixture-of-Experts Framework for Multi-Type Iatrogenic Drug Interaction Prediction

2026-06-12T11:07:08+07:00

Pharmacovigilance involves preventing adverse effects in patients through the prevention, detection, and evaluation of these effects. Prescription errors, incorrect dosages, and drug interactions cause iatrogenic drug effects. There are two main types of drug interactions: between two drugs (DDI) and between a protein and a drug (DTI). The multitude of molecules and treatments available today makes it difficult to prevent their effects. Deep learning (DL) appears to be a solution to this problem. Current work in DL focuses on a single type of interaction and does not generalize to other types. We propose a multitask model that predicts both types of interactions (DDI and DTI). It deploys two mixture-of-experts (MoE) blocks for each task. This mechanism avoids interference between representation spaces. One module for drugs uses convolutional neural networks (CNNs). A second module for proteins learns multi-granularity patterns. The model test yields the following results for the DTI task: accuracy (0.952), precision (0.959), F1 (0.952), recall (0.944), AUC (0.987), and AUPR (0.987). The DDI task yields accuracy (0.980), precision (0.966), F1 (0.980), recall (0.995), AUC (0.997), and AUPR (0.995).

A Decentralized Identity and Trust Management Platform Using Blockchain for Cross-Domain Authentication

2026-06-12T14:01:09+07:00

As digital ecosystems expand, secure and interoperable identity management across organizational boundaries has become increasingly important. This paper presents a blockchain-based platform for decentralized identity and trust management to support cross-domain authentication and authorization among autonomous organizations, such as government agencies and academic institutions. The proposed platform employs a consortium blockchain as a tamper-resistant credential and policy repository, enabling each organization to administer its own credentials while supporting verifiable identity sharing across domains. On-ledger trust relationships and authorization policies allow trusted interactions without relying on centralized identity authorities or pre-established bilateral agreements. A prototype was implemented using Hyperledger Fabric and evaluated in a multi-domain setting. The results demonstrate correct authentication behavior, sub-second authentication latency, measurable transaction throughput, and effective revocation propagation. Additional experiments under multi-domain and concurrent authentication workloads show that the platform preserves consistent authentication outcomes while maintaining latency within practical bounds. The proposed approach can be applied to multi-institutional environments, such as inter-university digital services, cross-agency e-government systems, and collaborative research infrastructures, where secure identity sharing and cross-domain access control are required.

Reinforcement Learning for Photometric Tuning Improves Ovarian Lesion Segmentation

2026-06-18T09:22:34+07:00

Image segmentation is a fundamental task in computer vision and has become an essential component of many real-world applications, particularly in medical image analysis. In ovarian tumor diagnosis, accurate ultrasound image segmentation can assist clinicians by reducing the time required for lesion delineation and facilitating faster treatment planning. However, ultrasound imaging remains challenging because images typically exhibit low contrast, speckle noise, and poorly defined lesion boundaries. These characteristics substantially degrade the performance of automated segmentation models, and reduce the accuracy of tumor localization. In this study, we introduce a reinforcement learning (RL)based method designed to improve image contrast, and effectively mitigate this problem. We apply a Proximal Policy Optimization (PPO) Agent that applies small, structure-preserving photometric transformations (contrast/gamma adjustment, noise reduction and contrast-limited adaptive histogram equalization) to the input image and a pretrained model (U-Net) to evaluate the output results. The Agent receives a distinguishable reward based on the improvement in the soft-Dice Similarity Coefficient (S-DSC) from one step to the next, with penalties for unnecessary actions and an explicit stopping action. This method improves S-DSC over the baseline U-Net without retraining or architectural changes to the segmentation set. The proposed method consistently improved segmentation performance across all evaluated cases. The most substantial improvement was observed in challenging ultrasound images, where the Dice score increased by up to 78%. These results demonstrate that adaptive, learned photometric adjustments can effectively enhance inference-time segmentation without requiring retraining or architectural modifications to the segmentation model. The approach is designed to be architecture-independent, easy to integrate with frozen segmenters, and to maintain full interpretability across step-by-step action traces.

THAISE: A Web Security Assessment Framework and Empirical Study of 126 Thai Higher Education Institutions

2026-06-12T09:40:18+07:00

The security of university websites is a vital but often overlooked issue, especially in developing countries. This work introduces THAISE (Thai Higher Education Institutions Security Evaluator), a system created to measure and score the web security levels of Thai universities. THAISE assesses seven key areas, including TLS configuration, certificate validity, and security headers such as HTTP Strict Transport Security (HSTS) and Content Security Policy (CSP). These factors are combined into a single Web Security Score (WSS). We used this framework to evaluate all 126 institutions under the Ministry of Higher Education, Science, Research and Innovation (MHESI). Our findings established a national baseline with an average score of 55.9. To ensure accuracy, we compared our results with industry tools like SecurityHeaders.com and Qualys SSL Labs, which showed consistent alignment. The results highlighted that Content Security Policy (CSP) is the most significant weakness, as 92.1% of universities failed to use it. Interestingly, exploratory analysis did not identify statistically significant associations between WSS and institutional budget per student, enrolment size, international ranking status, or Reinventing University group classification in the analyzed dataset. These findings suggest that web security posture may depend less on institutional resources or prestige than on deliberate configuration practices, though further research is needed to confirm this interpretation. These insights offer important directions for national cybersecurity policies in Thai higher education.

Enhancing Zero-Day and Known Attacks Detection using Hybrid-Optimized Stacking Ensemble Classifier

2026-06-12T11:38:34+07:00

Zero-day and malicious attacks are still a research challenge for host system security. Zero-day attacks targeting the host systems digital data. To address this problem, the proposed two-stage multi-layered framework detects both zero-day and known attacks. The first stage uses a correlation-based filtering approach. Followed by feature selection using the Boruta algorithm and then the autoencoder (BAE) anomaly detection module. The second stage develops a stacking ensemble classifier (SEC) and optimizes it using the Swarm-Bayesian Hybrid Optimization (SBHO) technique. The combined approach forms the BAE-SEC-SBHO model. The model has been evaluated on the UNSW Windows ToN 2020 dataset, which includes host behavioral features like process, disk, and memory activity. The suggested pipeline reduces false-negative rates and improves recall rates compared to traditional machine learning and deep learning models. The performance and robustness of the proposed model are validated through an ablation study and a Wilcoxon signed-rank test for host-based intrusion detection.

Robust Violence Detection in Heterogeneous Surveillance Videos using a Dual-Branch Spatiotemporal Network

2026-05-28T09:19:18+07:00

Automatic detection of violent events in surveillance video is a critical component of modern public safety systems. In contrast, existing methods often show limited generalization, performing well on individual datasets but degrading under domain shifts in real-world environments. To address this generalization challenge, this paper proposes an Asymmetric Dual- Branch Network trained on a large-scale Multi-Source Dataset integrating nine heterogeneous benchmark datasets. The proposed architecture combines a motion branch based on ResNet-18 trained from scratch to capture motion-specific representations with a spatial branch employing an X3D backbone selected through comparative evaluation of four state-of-the-art architectures, to learn complementary spatial semantic representations. This asymmetric design effectively balances detection accuracy and real-time performance while improving robustness across heterogeneous scenarios. Extensive experiments using 5-fold cross-validation demonstrate stable performance across multiple datasets. Notably, the proposed framework achieves state-of-the-art results on the Bus-violence dataset (89.07%) and RLVS dataset (96.65%), while maintaining a high processing rate of 76.61 FPS. These empirical results, including validated performance on the unseen Kranok-NV dataset, confirm that the asymmetric integration of decoupled features offers a practical and scalable solution for real-time urban surveillance in uncontrolled environments.